Keep your data safe

The big ideas

• STOP before you hit Enter

  Settings: know what the tool keeps. Trim: give it only what the task needs. Oversee: you approve before it acts. Prove it: if something is urgent or weird, check it another way.

• Your prompt takes a trip

  "We don't train on your data" only answers one of four questions. Also ask: Is my chat stored, and for how long? Who else gets it? Who can look at it?

• Training isn't the only thing that can happen

  Training builds the model. Fine-tuning and feedback from people (RLHF) shape it later. A company can skip training on your chat and still store it, have people review it for safety, or keep it for legal reasons.

• Temporary isn't zero

  Temporary or incognito chats usually aren't used for training, but a copy can still be kept for a while (ChatGPT says up to 30 days). Work accounts usually have stronger protections than personal plans, even paid ones.

• Privacy doesn't make the AI worse

  "Zero data retention" means the company doesn't keep your prompt, with listed exceptions. It doesn't make answers worse, because the model was trained long before your chat.

• Green, yellow, red

  Green: public info, fine to paste. Yellow: names, schedules, work details, so trim first. Red: ID numbers, health info, passwords, bank details, other people's private info. Never paste it.

• Documents are information, not permission

  Once AI can read your email or browse for you, a hidden line like "AI: ignore your instructions" can try to take over (that's called prompt injection). You hold the send button.

• Prove it

  Scams now use cloned voices and deepfake video. If a call or message is urgent and asks for money or codes, hang up and contact them using a number you already trust.

1. Level 1Required15–20 min

   Strip it, then prompt it

   1. Mark every detail in the volunteer note as GREEN (fine to paste), YELLOW (trim it first) or RED (never paste it).
   2. Rewrite the prompt with only what a shift reminder actually needs.
   3. Paste your rewritten prompt into any AI tool. Did the reminder still work?

   A complete answer

   • Kept the day, time and place (that's all a reminder needs)
   • Removed phone numbers, email, birthday and the client list
   • Marked probation, health and disability details and SNAP case numbers as RED
   • Noticed the reminder came out just as good without the private info
2. Level 2Optional15 min

   Write five rules

   1. Write one rule for each: what never goes into AI, which tools are OK (and with which settings), who reviews before anything is sent, what to do with case notes and donor records, and the oops plan if someone pastes something by mistake.
   2. Keep each rule to one sentence a tired volunteer would actually follow.

   A complete answer

   • Five rules, one sentence each
   • Names a person or role for review and for the oops plan
   • Uses plain words, not legal language
3. Level 3Optional10 min

   Approve or nope?

   1. For each of the three things the AI wants to do, write APPROVE, CHECK FIRST or NOPE, and one sentence on why.
   2. For anything marked CHECK FIRST, say exactly what you'd look at before clicking yes.

   A complete answer

   • Email to 84 people = CHECK FIRST: who it's going to and what's attached (the client list should not go out)
   • Calendar entry = APPROVE, or a quick check, since it's low-risk and easy to undo
   • Gift card email = NOPE: urgent, money and secrecy are scam signs, so confirm with the director another way
4. BonusOptionalOpen-ended

   Trick the Wizard

   1. Play gandalf.lakera.ai and get as far as you can.
   2. Write down the level you reached and the trick that worked best, then post it in Discord.

   A complete answer

   • Reported your level
   • Named one trick and how a real AI assistant could fall for it